Journal of Computational and Applied Mathematics 237 (2013) 529–541

Contents lists available at SciVerse ScienceDirect

Journal of Computational and Applied

Mathematics
journal homepage: www.elsevier.com/locate/cam

An algorithm for computing a Padé approximant with minimal

degree denominator
O.L. Ibryaeva ∗ , V.M. Adukov
South Ural State University, Lenin avenue 76, Chelyabinsk, Russia

article info abstract

Article history: In this paper, a new definition of a reduced Padé approximant and an algorithm for its
Received 24 January 2012 computation are proposed. Our approach is based on the investigation of the kernel
Received in revised form 10 June 2012 structure of the Toeplitz matrix. It is shown that the reduced Padé approximant always
has nice properties which the classical Padé approximant possesses only in the normal
Keywords: case. The new algorithm allows us to avoid the appearance of Froissart doublets induced
Padé approximant
by computer roundoff in the non-normal Padé table.
Toeplitz matrix
Padé–Laplace method
© 2012 Elsevier B.V. All rights reserved.
Froissart doublets

1. Introduction

Padé approximants are locally the best rational approximations that can easily be constructed from the coefficients of
a given power series. They are closely related to continued fractions, orthogonal polynomials, and Gaussian quadrature
methods [1]. They have been widely used in various problems of mathematics, physics, and engineering due to their property
to effectively solve the problem of analytic continuation of the series beyond its disc of convergence [2].
There are many methods available to compute Padé approximants [3]. Some of them are implemented in computer
algebra systems such as Maple and Mathematica, and their built-in utilities are frequently used in applied problems. For
calculation with floating-point numbers in Maple, an algorithm due to Geddes [4] based on fraction-free symmetric Gaussian
elimination is used. Recursive algorithms for computing Padé approximants are also widely distributed. At first, they allowed
one to find Padé approximants in the case of a normal Padé table, but later some of them were generalized to the non-normal
case; see, for example, [5].
It is well known that the Padé table of a rational function R(z ) is always non-normal since it contains an infinite singular
block whose elements are identical to R(z ). Actually the entries inside that block usually differ from the rational function R(z )
through the appearance of supplementary common roots in the numerators and the denominators, but, after reducing the
common factors, we get the rational function R(z ) (the Padé fraction in Gragg’s terminology [6]). In any practical calculations
(because of the computer roundoff and the noise presence in the input data from which Padé approximants are constructed),
the paired roots in the numerators and the denominators will not be rigorously equal. This phenomenon of «pairing» of such
zeros and poles was named the Froissart phenomenon, and the pairs are known as Froissart doublets [7]. For example, their
appearance is inevitable in signal processing using the Padé–Laplace method [8]. In order to identify doublets, several Padé
approximants (besides the desired one) are usually calculated. This requires additional coefficients of the Taylor series, and
is undesirable, since the coefficients are often computed numerically and Padé approximants are known to be very sensitive
to errors in the coefficients.
The main purpose of the present paper is to propose a new algorithm for computing Padé approximants. This algorithm
finds the denominator with the minimal degree among all the denominators of the Padé approximant and is based on

∗ Corresponding author.
E-mail address: oli@6v6power.ru (O.L. Ibryaeva).

0377-0427/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.cam.2012.06.022
530 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541

the results concerning the kernel structure of Toeplitz matrices [9]. In the singular block case it allows one to avoid the
appearance of Froissart doublets induced by computer roundoff.
Recently, a similar algorithm for computing Padé approximants was proposed in [10]. The authors used the method of
hopping across a square singular block of the Padé table to find its upper-left corner. Having found the rank of a matrix, they
reduce the numerator and denominator degrees diagonally and move to another position in the singular block. There may
be several such hops across the singular blocks before we get to the final position in the upper-left corner corresponding to
the minimal degree denominator. In our algorithm, we find this denominator by another method, reducing the degree in
the denominator alone. The use of singular value decomposition (SVD) for computing numerical ranks (in the present paper
and in [10]) realizes a regularization of the ill-posed Padé approximant problem and leads to stable algorithms.
Note that the concept of minimality of polynomial degrees is thoroughly studied in the theory of matrix Padé
approximant. This problem is much more complicated, since in the matrix case the concept of minimum degrees and
irreducible polynomials loses the properties of uniqueness [11]. In some cases, a rational matrix function can be represented
in irreducible form by different pairs of polynomials with different pairs of minimal degrees (in some sense). In the recent
paper [12], the authors proposed rational approximations of a matrix formal power series, with both matrix polynomials,
numerator and denominator, satisfying three conditions: (a) minimum row degrees for the numerator and denominator,
(b) an invertible denominator at the origin, and (c) canonical representation (without free parameters).
The paper is organized as follows. In Section 2, some definitions and examples are given. These examples show that
in order to avoid the appearance of Froissart doublets we should study the kernel of the Toeplitz matrix which gives us
denominators of the Padé approximant. This is done in Section 3, which contains the main result on the parameterization
of the set of all the denominators of the Padé approximant. This result allows us to give the definition of the modified
Padé approximant in Section 4 and to establish its properties. In Section 5, we propose our algorithm for Padé approximant
computation and provide some examples in order to show that it allows us to avoid Froissart doublets induced by computer
roundoff.

2. Preliminaries

This section contains some definitions and numerical experiments obtained in Maple and Mathematica. We have chosen
these packages because they are the most widespread symbolic computation systems used in research and applications.
The examples illustrate how Froissart doublets appear or do not appear when we deal with the singular block of a Padé
table for a rational function. Note that our new algorithm for computing Padé approximants will be proposed in Section 5.
We will implement it in the Maple system and solve some of the examples again in order to show that the new method
allows us to avoid Froissart doublets induced by computer roundoff.
We start with the classical definition of Padé–Frobenius approximants.

Definition 2.1 (Padé–Frobenius). Let f (z ) be a (formal) power series f (z ) = ck z k , ck ∈ C. The (m, n) Padé approximant
∞
k=0
Pm,n (z )
corresponding to f (z ) is the rational function fm,n (z ) = Qm,n (z )
, where Pm,n (z ) and Qm,n (z ) are polynomials in z such that

1. Qm,n (z ) ̸≡ 0, deg Qm,n (z ) ≤ n, deg Pm,n (z ) ≤ m,

2. f (z )Qm,n (z ) − Pm,n (z ) = rm+n+1 z m+n+1 + rm+n+2 z m+n+2 + · · ·.

Note that, in Gragg’s paper [6], the Padé–Frobenius approximant is called a Padé form. The Padé form in which the
denominator and numerator are coprime is called the Padé fraction in [6].
In a similar way, we can define the Padé–Frobenius approximant at the point z = a for the series k=0 ck (z − a)k . For
∞
the sake of brevity, we will call it the Padé approximant. Throughout the paper, we will use Padé approximants at the point
z = 0.
Obviously, the coefficients q0 , . . . , qn of the denominator Qm,n (z ) can be obtained by solving the following homogeneous
system of linear equations with a Toeplitz n × (n + 1) matrix:

cm+1 cm ··· cm−n+1 q0 0

    
cm+2 cm+1 ··· cm−n+2  q1  0
. .. .. .  =  . . (1)
. .   .
. . . . .
cm+n cm+n−1 ··· cm qn 0

Here and below we assume that ck = 0 if k < 0. From Condition 2 it follows that, with the q0 , . . . , qn available, the
coefficients p0 , . . . , pm of the numerator Pm,n (z ) can be obtained from the formula

p0 c0 0 ··· 0 q0
    
p 1  c1 c0 ··· 0  q1 
.  = . .. .. . . (2)
.  . . 
. . . . .
pm cm cm−1 ··· cm−n qn
 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541 531

Thus, the (m, n) Padé approximant fm,n (z ) always exists. Generally the polynomial Qm,n (z ) (and hence Pm,n (z )) is not
unique, since the rank of the matrix of the system (1) can be less than n. Nevertheless, it is easy to show that the rational
function fm,n (z ) is unique. Usually these rational functions are arranged in a double-entry table known as the Padé table
fm,n (z ) m,n=0,1,... corresponding to f (z ). The Padé approximant fm,n (z ) occupies the nth line and the mth column of the
table. It is well known that identical Padé approximants can occur only in square blocks of the table. If the Padé table does
not contain such blocks, it is said to be normal; otherwise, it is called non-normal.
Provided that there is denominator Qm,n (z ) such that Qm,n (0) ̸= 0, we can rewrite Condition 2 as follows:

Pm,n (z )
f (z ) − = Rm+n+1 z m+n+1 + Rm+n+2 z m+n+2 + · · · . (3)
Qm,n (z )
Pm,n (z )
In this case, the rational function Qm,n (z )
is called a Padé–Baker approximant. Note that the Padé–Baker approximant does
not always exist.
We now come to the second and the main part of this section, where we consider how Padé approximants can be obtained
in such popular mathematical packages as Maple, Mathematica, and Matlab.
The built-in utility pade in Matlab approximates time delays by rational linear-time invariant models, i.e., actually it
finds only diagonal Padé approximants for an exponential function by explicit formulas available for this case [13]. Maple
and Mathematica compute Padé approximants in the general case, and in the examples below we will use their functions
pade(f , z = a, [m, n]) (in Maple) and Pade Approximant [expr , z , a, m, n] (in Mathematica), which compute the (m, n) Padé
approximant of the function f (z ) at the point z = a.
To compare the results with the approximated rational functions we will find the zeros and poles of the approximants.
(z +1)(z −2)
Example 2.1. Let us find by Maple and Mathematica the diagonal (2, 2) Padé approximant f2,2 for f (z ) = (z +2.1)(z −1) at the
point z = 0 via the following commands.
In Maple, we have
(z +1)·(z −2)
f := (z +2.1)·(z −1) :
with(numapprox): p := pade(f , z = 0, [2, 2]);

0.9523809523 + 0.4761904764z − 0.4761904759z 2

0.9999999999 − 0.5238095237z − 0.4761904759z 2
fsolve(numer (p), z , complex);
−1.000000000, 2.000000001
fsolve(denom(p), z , complex);
−2.100000001, 1.000000000.
In Mathematica, we have
In[1] := f := (z + 1) ∗ (z − 2)/((z + 2.1) ∗ (z − 1));
In[2] := P := Pade Approximant [f , {z , 0, {2, 2}}]; P
0.952381 + 0.47619z − 0.47619z 2
Out [2] :=
1.000000000000000 − 0.52381z − 0.47619z 2
In[3] := Roots[Numerator [P ] == 0, z ]
Out [3] := z == −1. ∥ z == 2.
In[4] := Roots[Denominator [P ] == 0, z ]
Out [4] := z == −2.1 ∥ z == 1.

As can be seen, the obtained approximant f2,2 is identical to the function f (z ) since they have the same zeros and poles.
Note that in this case the kernel dimension of the matrix of system (1) is equal to 1, i.e., the denominator of the Padé
approximant is unique.
The following example shows that f3,3 is again identical to the function f (z ) (both in Maple and in Mathematica), though
the kernel is multidimensional.
(z +1)(z −2)
Example 2.2. Let us find the (3, 3) Padé approximant for the function f (z ) = (z +2.1)(z −1) at the point z = 0.
In Maple, we have
p := pade(f , z = 0, [3, 3]);

0.9523809523 + 0.4761904764z − 0.4761904759z 2

0.9999999999 − 0.5238095237z − 0.4761904759z 2
532 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541

fsolve(numer (p), z , complex);

−1.000000000, 2.000000001

fsolve(denom(p), z , complex);

−2.100000001, 1.000000000.

In Mathematica, we have

In[1] := P := Pade Approximant [f , {z , 0, {3, 3}}]; P

0.952381 + 0.47619z − 0.47619z 2 + 0.z 3
Out [1] :=
1.000000000000000 − 0.52381z − 0.47619z 2 + 0.z 3
In[2] := Roots[Numerator [P ] == 0, z ]
Out [2] := z == −1. ∥ z == 2.
In[3] := Roots[Denominator [P ] == 0, z ]
Out [3] := z == −2.1 ∥ z == 1.

However, approximant f4,4 does not coincide with f (z ) because of the appearance of supplementary roots of the
denominator and the numerator.
(z +1)(z −2)
Example 2.3. Let us find the (4, 4) Padé approximant for the function f (z ) = (z +2.1)(z −1) at the point z = 0.
In Maple, we have
p := pade(f , z = 0, [4, 4]);

0.9523809524 + 0.9750566894z − 0.2267573697z 2 − 0.2494331066z 3

1. − 0.7505668935z 2 − 0.2494331066z 3
fsolve(numer (p), z , complex);

−1.909090909, −0.9999999999, 2.000000000

fsolve(denom(p), z , complex);

−2.100000001, −1.909090908, 1.000000000.

In Mathematica, we have

In[1] := P := Pade Approximant [f , {z , 0, {4, 4}}]; P

0.952381 + 1.11851z − 0.155032z 2 − 0.321159z 3 − 3.33067 × 10−16 z 4
Out [1] :=
1.000000000000000 + 0.150624z − 0.829465z 2 − 0.321159z 3 + 0.z 4
In[2] := Roots[Numerator [P ] == 0, z ]
Out [2] := z == −9.64247 × 1014 ∥ z == −1.48273 ∥ z == −1. ∥ z == 2.
In[3] := Roots[Denominator [P ] == 0, z ]
Out [3] := z == −2.1 ∥ z == −1.48273 ∥ z == 1.

Thus, in Example 2.3, Froissart doublets induced by computer roundoff appear. Note that the doublets obtained in Maple
and Mathematica are different (they are given in bold). The root − 9.64247 × 1014 of the numerator (obtained in Mathematica)
is caused by its small leading coefficient −3.33067 × 10−16 .
Note that for f3,3 and f4,4 the denominator is not unique and, as we have seen, Maple and Mathematica may choose not
the best one.
Let us take another rational function and find its approximation in its infinite singular block. As it can be seen below,
Maple does not produce Froissart doublets, but they appear when using Mathematica.

1.01
Example 2.4. Let us find the (2, 3) Padé approximant for the function f (z ) = (z +2z +
)(z −2.01)
at the point z = 0.
In Maple, we have
+1.01
f := (z +2z)·(z −2.01)
:
with(numapprox) : p := pade(f , z , [2, 3]);
−0.2512437812 − 0.2487562190z
1.000000000 + 0.002487561992z − 0.2487562188z 2
 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541 533

fsolve(numer (p), z , complex);

−1.010000000
fsolve(denom(p), z , complex);
−2.000000001, 2.010000000.
In Mathematica, we have
In[1] := f := (z + 1.01)/((z + 2) ∗ (z − 2.01));
In[2] := P := PadeApproximant [f , {z , 0, {2, 3}}]; P
−0.251244 − 0.122069z + 0.125433z 2
Out [2] :=
1.000000000000000 − 0.501752z − 0.250011z 2 + 0.125433z 3
In[3] := Roots[Numerator [P ] == 0, z ]
Out [3] := z == −1.01 ∥ z == 1.98318
In[4] := Roots[Denominator [P ] == 0, z ]
Out [4] := z == −2. ∥ z == 1.98318 ∥ z == 2.01.

The following conclusions may be drawn from the examples. When the kernel dimension of the matrix of system
(1) is equal to 1 (as in Example 2.1), we have one denominator of the Padé approximant, and Froissart doublets cannot
appear. But they can sometimes appear when the kernel dimension is greater than 1. If an improper denominator is chosen,
then supplementary (also called artificial) roots appear. These roots may have corresponding pairs among the roots of the
numerator, and the cancelation of the common factors reduces the Padé approximant to the reduced form. But, in the
presence of computer roundoff, the paired roots in the numerator and denominator will not be rigorously equal, and this
can cause the appearance of Froissart doublets (as in Examples 2.3 and 2.4). Moreover, computer roundoff may produce
artificial roots with large absolute value (as in Example 2.3), since the vanishing leading coefficients are small but not equal
to zero.
To avoid these problems, we should choose the denominator with the minimal degree. As will be shown further, this
denominator always exists. In order to find it, we have to study the kernel structure of the Toeplitz matrix of system (1).
This will be done in the following section.

3. Parameterization of the denominator set

In this section, we study the structure of the set of all denominators for an (m, n) Padé approximant, i.e., the structure of
the kernel ker Tm+1 of the Toeplitz matrix
Tm+1 = ∥ci−j ∥ i=m+1,...,m+n
j=0,1,...,n

from system (1).

First, let us prove that the minimal degree denominator exists, and establish its properties.

Proposition 3.1. There exists the denominator Qm0 ,n (z ) with the minimal degree among all the denominators Qm,n (z ). This
denominator is unique up to a constant factor.
,n (z ) be the numerator of the Padé approximant corresponding to Qm,n (z ). Then the polynomials Pm,n (z ) and Qm,n (z ) do
0 0 0 0
Let Pm
not have common non-zero roots.

Proof. Since all the polynomials Qm,n (z ) satisfy the condition deg Qm,n (z ) ≤ n, there exists a denominator with the minimal
degree d.
Suppose that there are two denominators with degree d:
Qm0 ,n (z ) = Bd z d + · · · + B0 , Qm0 ,n (z ) = 
 B0 ,
Bd z d + · · · +  Bd ̸= 0, 
Bd ̸= 0.

Let us introduce Q (z ) = 
Bd Qm0 ,n (z ) − Bd 
Qm0 ,n (z ). This is the denominator of the Padé approximant with degree less than
d. This is possible if and only if Q (z ) ≡ 0. Thus, Qm0 ,n (z ) and 
Qm0 ,n (z ) are linearly dependent. The uniqueness is proved.
Let Pm,n (z ) be the numerator corresponding to Qm,n (z ). Let us suppose that z0 is the common non-zero root of Pm
0 0 0
,n (z )
and Qm0 ,n (z ):

,n (z ) = (z − z0 )Pm,n (z ), Qm0 ,n (z ) = (z − z0 )
Qm0 ,n (z ).
0 0
Pm
By Definition 2.1, we have
1
f (z )
Qm0 ,n (z ) − 0
,n (z ) = O z m+n+1 = O z m+n+1 .
   
Pm
(z − z0 )
Qm0 ,n (z ) is the denominator with degree less than d. We have a contradiction, since d is the minimal degree.
Then  
534 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541

Now we describe the denominator set {Qm,n (z )}, i.e., the kernel structure of Tm+1 . This result will be used to construct
the algorithm for obtaining the minimal degree denominator.
To find any denominator of the Padé approximant, we need the sequence of the Taylor coefficients cm−n+1 , . . . , cm+n of
f (z ). The sequence consists of the entries of matrix Tm+1 . Let us introduce notions of indices and essential polynomials for
cm−n+1 , . . . , cm+n . The notions were given in more general situation in [9]. Here, we formulate them specially for our case.
It is natural to include Tm+1 in the family of Toeplitz matrices

ck ck−1 ··· cm−n+1

 
 ck+1 ck ··· cm−n+2 
, m − n + 1 ≤ k ≤ m + n,
 ..
Tk =  .. ..
.
.. (4)
. . .

cm+n cm+n−1 ··· c2m−k+1
which are constructed from the elements of cm−n+1 , . . . , cm+n .
Consider the sequence of spaces ker Tk . It is more convenient to deal not with vectors Q = (q0 , q1 , . . . , qk−m+n−1 )t ∈
ker Tk but with their generating polynomials Q (z ) = q0 + q1 z + · · · + q k−m+n−1 z k−m+n−1 . Instead of the spaces ker Tk we
will use the isomorphic spaces Nk consisting of the generating polynomials.
To do this, we introduce a linear functional σ by the formula
σ {z j } = c−j , −m − n ≤ j ≤ −m + n − 1.
In the theory of orthogonal polynomials this functional is called the Stieltjes functional.
Denote by Nk (m − n + 1 ≤ k ≤ m + n) the space of polynomials Q (z ) with formal degree k − m + n − 1 satisfying the
orthogonality conditions:

σ z −i Q (z ) = 0, i = k, k + 1, . . . , m + n.
 
(5)
It is easily seen that Nk is the space of generating polynomials of vectors in ker Tk . For convenience, we put Nm−n = 0
and denote by Nm+n+1 the (2n + 1)-dimensional space of all polynomials with formal degree 2n.
Let dk be the dimension of the space Nk , and let ∆k = dk − dk−1 (m − n + 1 ≤ k ≤ m + n + 1). The following fact is
crucial for our further considerations.

Proposition 3.2. For any non-zero sequence cm−n+1 , . . . , cm+n , the inequalities
0 = ∆m−n+1 ≤ ∆m−n+2 ≤ · · · ≤ ∆m+n ≤ ∆m+n+1 = 2 (6)
are fulfilled.
Proof. It follows from Definition 2.1 that Nk , z Nk are subspaces of Nk+1 (m − n ≤ k ≤ m + n), and that
Nk ∩ z Nk = z Nk−1 .
Hence, by the Grassman formula, we have

dim Nk + z Nk = 2dk − dk−1 .

 
(7)
Let hk+1 be the dimension of any complement Hk+1 of the subspace Nk + z Nk in the whole space Nk+1 . From (7), we
have
hk+1 = ∆k+1 − ∆k , (8)
i.e., ∆k+1 ≥ ∆k . By definition, dm−n = 0, and dm−n+1 is also equal to zero. Hence, ∆m−n+1 = 0. In a similar manner we can
prove that ∆m+n+1 = 2. 
It follows from inequalities (6) that there exist integers µ1 ≤ µ2 such that

∆m−n+1 = · · · = ∆µ1 = 0,
∆µ1 +1 = · · · = ∆µ2 = 1, (9)
∆µ2 +1 = · · · = ∆m+n+1 = 2.
If the second row in these relations is absent, we assume that µ1 = µ2 . In our case, µ1 < µ2 , as will be shown later.

Definition 3.1. The integers µ1 , µ2 defined in (9) will be called the essential indices (briefly, indices) of the sequence
cm−n+1 , . . . , cm+n .

Proposition 3.3. Let ~ = rank Tm . Then the indices µ1 , µ2 are found by the following formulas:
µ1 = m − n + ~, µ2 = m + n − ~ + 1.
m+n+1
Proof. It follows from the definition of ∆k that k=m−n+1 ∆k = dm+n+1 − dm−n = 2n + 1.
 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541 535

On the other hand, from the relations (9), we have

m+n+1

∆k = 1 · (µ2 − µ1 ) + 2 · (m + n + 1 − µ2 ).
k=m−n+1

Hence, µ1 + µ2 = 2m + 1. This means

m that µ1 ≤ m < µ2 .
In a similar manner, we obtain k=m−n+1 ∆k = dm = n − ~ and k=m−n+1 ∆k = m − µ1 . Thus, µ1 = m − n + ~ .
m
Since µ1 + µ2 = 2m + 1, we get µ2 = m + n − ~ + 1. 
Now, we can describe the structure of the kernels of the matrices Tk . It follows from (8) and (9) that hk+1 := dim Hk+1 ̸= 0
iff k = µj (j = 1, 2). In that case, hk+1 = 1. Therefore, for k ̸= µj ,
Nk+1 = Nk + z Nk , (10)
and, for k = µj ,

Nk+1 = Nk + z Nk +̇Hk+1 .
 
(11)

Definition 3.2. Any polynomial Qj (z ) that forms a basis for the one-dimensional complement Hµj +1 will be called the
essential polynomial of the sequence cm−n+1 , . . . , cm+n corresponding to the index µj , j = 1, 2.
It can be shown (see [9, Theorem 4.1]) that integers µ1 , µ2 are the indices, and that polynomials Q1 (z ) ∈ Nµ1 +1 , Q2 (z ) ∈
Nµ2 +1 are the essential polynomials iff
σ0 := σ {z −m−n−1 [Q2 (0)Q1 (z ) − Q1 (0)Q2 (z )]} ̸= 0. (12)
In the following theorem, the structure of the kernels of matrices Tk (m − n + 1 ≤ k ≤ m + n) is described in terms of
the indices and the essential polynomials.

Proposition 3.4. Let µ1 , µ2 be the indices and let Q1 (z ), Q2 (z ) be the essential polynomials of the sequence cm−n+1 , . . . , cm+n .
Then
0, m − n + 1 ≤ k ≤ µ1 ,

Nk = {q1 (z )Q1 (z )}, µ1 + 1 ≤ k ≤ µ2 ,
{q1 (z )Q1 (z ) + q2 (z )Q2 (z )}, µ 2 + 1 ≤ k ≤ m + n,
where q1 (z ), q2 (z ) are arbitrary polynomials of formal degree k − µj − 1.
Proof. From (9), we have dk = 0 for k ∈ [m − n + 1, µ1 ), i.e., Nk = 0.
Let µ1 + 1 ≤ k ≤ µ2 . It follows from (10) and (11) that the polynomials

R1 (z ), zR1 (z ), . . . , z k−µ1 −1 R1 (z )
 
(13)

k to k − µ1 .
generate the space Nk . The number of these polynomials is equal
From the definition of ∆j and relations (9), we have dk = j=m−n+1 ∆j = k − µ1 .
Thus, the number of polynomials in (13) is equal to the dimension of the space Nk . Therefore, these polynomials form a
basis of Nk .
The case µ2 + 1 ≤ k ≤ m + n can be considered in a similar manner. 
Now we apply the previous proposition to the case k = m + 1 and obtain the main result of this section on the
parameterization of the denominator set.

Theorem 3.1. Let Qm,n (z ) be an arbitrary denominator of the (m, n) Padé approximant for the series f (z ) =
∞
k=0 ck z k . Form
the sequence {cm−n+1 , . . . , cm+n }, which is necessary to find Qm,n (z ).
Let µ1 be the first index, and let Q1 (z ) be the first essential polynomial of this sequence.
Then the denominator set is
Qm,n (z ) = {q1 (z )Q1 (z )} ,
 
(14)

where q1 (z ) is an arbitrary polynomial with formal degree m − µ1 . Thus, Q1 (z ) is the denominator Qm0 ,n (z ) with the minimal
degree.

Remark 3.1. It follows from Qm,n (z ) = q(z )Q1 (z ) that the denominator Qm,n (z ) such that Qm,n (0) ̸= 0 (Baker’s condition)
exists iff Q1 (0) ̸= 0.
Hence, the Padé approximant with the minimal degree denominator is the Padé–Baker approximant if the latter exists.

Remark 3.2. Our parameterization (14) is similar to Gragg’s parameterization in [6, Theorem 3.2, Item c]. However, to find
the minimal degree denominator in Gragg’s Theorem 3.2, we have to find any Padé form and reduce it by solving an ill-posed
problem of computing the greatest common divisor (GCD) of two polynomials. Our Theorem 3.1 gives a GCD-free method
for obtaining the minimal degree denominator Q1 (z ).
536 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541

4. Reduced Padé approximant and its properties

In this section, we use the parameterization in order to modify the definition of the Padé approximants.
Definition 2.1 ignores the non-uniqueness (in the general case) of a denominator of a Padé approximant. It does not
matter much for the normal Padé table, but in the singular case common factors of the denominator and the numerator
cannot be cancelled because of roundoff errors. This is one of reasons why Froissart doublets can appear.
The main aim of this section is to modify the classical definition, Definition 2.1. We would like a Padé approximant to
exist always, to have a unique denominator (up to a constant factor) and to coincide with the Padé–Baker approximant if
the latter exists.
Let us add the minimality requirement to Definition 2.1.

Definition 4.1. Let f (z ) be a (formal) power series f (z ) = ck z k , ck ∈ C. The (m, n) Padé approximant corresponding
∞
k =0
Pm,n (z )
to f (z ) is the rational function fm,n (z ) = Qm,n (z )
, where Pm,n (z ) and Qm,n (z ) are polynomials in z such that

1. Qm,n (z ) ̸≡ 0, deg Qm,n (z ) ≤ n, deg Pm,n (z ) ≤ m,

2. f (z )Qm,n (z ) − Pm,n (z ) = rm+n+1 z m+n+1 + rm+n+2 z m+n+2 + · · ·.
3. The polynomial Qm,n (z ) has the minimal degree among all polynomials satisfying 1, 2.

Theorem 3.1 on parameterization of the denominator set gives a constructive method for finding the denominator with
the minimal degree.
The next theorem shows that the Padé approximant has the desired properties.

Theorem 4.1. For any power series f (z ), the following statements are fulfilled.

1. Any (m, n) Padé approximant exists and is unique.

2. The denominator Qm,n (z ) of the Padé approximant is unique up to a constant factor, and Qm,n (z ) is the first essential polynomial
Q1 (z ) of the sequence {cm−n+1 , . . . , cm+n }.
3. Pm,n (z ) and Qm,n (z ) do not have common non-zero roots.
4. The Padé–Baker approximant exists iff Q1 (0) ̸= 0. The Padé approximant from Definition 4.1 is the Padé–Baker approximant
if the latter exists.
5. If Q1 (z ) has root z = 0 of order δm,n > 0, then

f (z ) − fm,n (z ) = Az m+n+1−δm,n + Bz m+n+2−δm,n + · · · , A ̸= 0.

(δm,n is called the deficiency index of the Padé approximant.)

Proof. Statements 1–4 evidently follow from Proposition 3.1, Theorem 3.1, and Remark 3.1.
We prove the last statement of the theorem. Let z = 0 be the root of order δm,n of Q1 (z ). Since Q1 (0) = 0, the number σ0
from formula (12) is

σ0 = −σ {z −m−n−1 Q1 (z )}Q2 (0) ̸= 0.

Here Q2 (z ) is the second essential polynomial.
Thus σ {z −m−n−1 Q1 (z )} ̸= 0.
It is easy to see that the number σ {z −m−n−1 Q1 (z )} is the coefficient at z m+n+1 in the power series f (z )Q1 (z ). As it is not
zero, we have

f (z )Q1 (z ) − P1 (z ) = A1 z m+n+1 + · · · , A1 = σ {z −m−n−1 Q1 (z )} ̸= 0.

Here, P1 (z ) is the numerator corresponding to the denominator Q1 (z ).

After dividing this equation by Q1 (z ), we get

f (z ) − fm,n (z ) = Az m+n+1−δm,n + · · · ,

where A ̸= 0.
This means that the deficiency index of the Padé approximant fm,n (z ) coincides with the multiplicity δm,n of the root
z = 0 of the first essential polynomial Q1 (z ). 

Since the numerator and the denominator of Padé approximant from Definition 4.1 do not have common non-zero roots,
we will call this approximant the reduced Padé approximant. Note that the numerator and the denominator of the reduced
Padé approximant cannot be coprime because of their common zero roots.
 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541 537

Remark 4.1. It is impossible to improve Definition 4.1 in such a way that the denominator and the numerator are always
coprime.
Pm,n (z )
0
It is easy to see that, if Qm,n (z ) = z δ Qm0 ,n (z ), then Pm,n (z ) = z δ Pm
0
,n (z ). Hence fm,n (z ) = . This rational function is
,n (z )
0
Qm
the Padé fraction in Gragg’s terminology. However, the polynomials (z ), (z ) cannot be considered as the numerator
0
Pm ,n Qm0 ,n
and the denominator of Padé approximant fm,n (z ), since the Frobenius Condition 2 is not fulfilled:

f (z )Qm0 ,n (z ) − Pm
0
,n (z ) = Az
m+n−δ+1
+ ···.

To determine δ (the deficiency index of fm,n (z )), we should know how many low-order coefficients of Q1 (z ) are equal to
zero. Note that vanishing coefficients of the denominator and the numerator can be non-zero because of roundoff errors. In
particular, the non-zero vanishing coefficients can cause the appearance of roots of the denominator and/or the numerator
with the great absolute value (as in Example 2.3).
Thus, we face the problem of finding vanishing coefficients of the denominator and the numerator. The results of
Theorem 4.1 allow us to solve this problem. This will be done in the following two theorems.
Recall that ~ = rank Tm , µ1 = m − n + ~ , matrix Tµ1 +1 has size (2n − ~) × (~ + 1), and rank Tµ1 +1 = ~ . The vector
which is a basis for one-dimensional space ker Tµ1 +1 gives coefficients of the first essential polynomial Q1 (z ) which is the
minimal degree denominator of the Padé approximant.
(k)
Further, we will denote by Tµ1 +1 the matrix which is obtained from the matrix Tµ1 +1 by deletion of the kth column,
k = 1, . . . , ~ + 1.
(k+1)
Theorem 4.2. Let Q1 (z ) = q0 + q1 z + · · · + q~ z ~ be the denominator with the minimal degree. Then qk = 0 iff rank Tµ1 +1 =
~ − 1, k = 0, . . . , ~ .
(k+1) (k+1) (k+1)
Proof. Let qk = 0. We have rank Tµ1 +1 ≤ ~ . Suppose rank Tµ1 +1 = ~ . Then Tµ1 +1 has a trivial kernel, i.e., q0 = q1 = · · · =
(k+1)
qk−1 = qk+1 = · · · = q~ = 0. Thus, Q1 (z ) ≡ 0. Since this is impossible, we have rank Tµ1 +1 ≤ ~ − 1.
(k+1) (k+1)
Assume that rank Tµ1 +1 < ~ − 1. Then Tµ1 +1 has a multidimensional kernel, which can be embedded in the kernel of
(k+1)
Tµ1 +1 in a natural way. But ker Tµ1 +1 is one dimensional; hence our assumption is false, and rank Tµ1 +1 = ~ − 1.
(k+1) (k+1)
On the other hand, if rank Tµ1 +1 = ~ − 1, then ker Tµ1 +1 is one dimensional, and after the natural embedding in one-
dimensional space ker Tµ1 +1 , we get qk = 0. 

Denote by Tµ1 +1 the matrix which is obtained by inserting the row (ck ck−1 . . . ck−~ ) at the beginning of the matrix
[k]

Tµ1 +1 , k = 0, . . . , m.

Theorem 4.3. Let P1 (z ) = p0 + p1 z + · · · + pm z m be the numerator corresponding to the denominator Q1 (z ). Then pk = 0 iff
rank Tµ1 +1 = ~, k = 0, . . . , m.
[k]

Proof. It follows from (2) that pk = ck q0 + ck−1 q1 + · · · + ck−~ q~ , k = 0, . . . , m. Hence,

q0 pk
   
 q1  0
[k]  =  . .
 ..
T µ 1 +1   .
. .
q~ 0

If pk = 0, then matrix Tµ1 +1 has a non-trivial kernel; hence rank Tµ1 +1 < ~ + 1. Since rank Tµ1 +1 = ~ , we have
[k] [k]

rank Tµ1 +1 = ~ .
[k]

On the other hand, if rank Tµ1 +1 = ~ then the inserted row (ck ck−1 . . . ck−~ ) is a linear combination of the rows of Tµ1 +1 .
[k]

Therefore, ck q0 + ck−1 q1 + · · · + ck−~ q~ = 0, i.e., pk = 0. 

From Theorems 4.2 and 4.3, we can obtain the multiplicities of z = 0, z = ∞ as the roots of Q1 (z ), P1 (z ), i.e., the
deficiency index of fm,n (z ) and the degrees of Q1 (z ), P1 (z ).
In particular, the following result on the deficiency index is now evident.

Corollary 4.1. The number δ is the deficiency index of fm,n (z ) iff

(1) (δ−1) (δ)
rank Tµ1 +1 = · · · = rank Tµ1 +1 = ~ − 1, rank Tµ1 +1 = ~.

We would like to end this section with the following remark. As is well known in the case of normal Padé approximant
(z )
fm,n (z ) = Qm,n (z ) , the degrees of Pm,n (z ) and Qm,n (z ) are equal to m and n, respectively, and δ = 0. Thus, in the normal
P
m,n
case, the degrees of the numerator and the denominator and the deficiency index δ are known. Our modified definition
(Definition 4.1) and the results of this section allow us to determine them not only in the normal case, but also in the non-
normal one.
538 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541

5. Algorithm

In this section, we present our algorithm for computing a reduced Padé approximant with the minimal degree
denominator. As we have seen in the previous section, in order to realize the algorithm, we have to find the rank of matrix
Tm , and the null space of Tµ1 +1 . Moreover, in order to delete the vanishing coefficients of the denominator and the numerator
(k)
we have to find the ranks of matrices Tµ1 +1 , k = 1, . . . , ~ + 1, and Tµ1 +1 , k = 0, . . . , m. In practice, due to rounding and
[k]

measuring errors and finite computer precision, the elements of these matrices are perturbed with error, so we have to
determine the rank and the null space of the original matrix from the perturbed matrix.
For computer calculations we need to define the numerical rank and the numerical null space of A. Recall these concepts
which are crucial for the algorithm.
The definition of the numerical rank was first given in [14]. We will use the simplified definition (as in [15,13, p. 72]).
The numerical ε rank of an M × N matrix A with respect to the threshold ε > 0 is defined as the smallest rank of all matrices
within a 2-norm distance ε of A. Namely,
rank (A, ε) = min{rank B :∥ A − B ∥2 ≤ ε}.
B

The numerical ε rank may be characterized in terms of the singular value decomposition (SVD). According to Stewart [16],
«singular value decomposition is the crème de la crème among rank-revealing decomposition».
Let us recall the definition of SVD. Every M × N complex matrix A can be represented in the form
A = UΣV H,
where U , V are unitary matrices (H means the Hermite conjugation), and Σ is an M × N matrix in which the upper N × N
block is a diagonal matrix with all entries real and sorted in descending order. The diagonal entries σ1 ≥ · · · ≥ σN of the
matrix Σ are called the singular values.
In the case of exact calculations, we have σ1 ≥ · · · ≥ σr > σr +1 = · · · = σN = 0, where r is the rank of A. In practice,
we have
σ1 ≥ · · · ≥ σr > ε ≥ σr +1 ≥ σN ≥ 0.
Let us denote Ar = U Σr V H with Σr = diag{σ1 , . . . , σr , 0, . . . , 0}; then ∥A − Ar ∥2 = σr +1 and rank (A, ε) = rank Ar = r [14].
Moreover, Ar is the nearest matrix to A (with respect to the 2-norm) with rank r. Therefore the null space of Ar is called the
numerical null space of A within ε . The null space of Ar is spanned by {vk+1 , . . . , vN }, where vk is the kth column of matrix
V . The ratio γ = σσr is called the numerical rank gap. The numerical rank of the matrix A may be estimated reliably in the
r +1
case of «well-defined numerical rank»; that is, when A has a single well-determined gap between large and small singular
values.
Thus, in terms of SVD, the numerical ε rank is the number of singular values greater than the given threshold ε . There
is no uniform threshold for all applications. The user must make a decision on the threshold ε based on the nature of the
application.
The rank(A,tol) function in Matlab returns the number of singular values of A that are larger than the threshold
(tolerance) tol. The default tolerance is max(M , N ) ∗ eps(norm(A)). Here, norm(A) indicates the Euclidean norm of A and
eps(norm(A)) is approximately 2.2 × 10−16 times norm(A). This choice is usually a good choice if the errors in the matrix
elements are due to computer arithmetic and if there is a sufficiently large gap in the singular values around this tolerance.
In [10], the authors treat a singular value as zero if it is less than tol ·∥c ∥2 , where c = (c0 , . . . , cm+n ) and tol is a relative
tolerance. The tolerance is applied also in the detection of zero coefficients in the numerator and the denominator. The
chosen value of tol in [10] is 10−14 . Our first version of the algorithm [17] used two independent parameters (tolerances)
for these two goals: (1) for determining if the singular value is zero and (2) for determining if the coefficient is zero. Now,
due to Theorems 4.2 and 4.3, we can choose only one parameter. In some applications the use of two tolerances may be
preferable.
Although SVD is the most widely used method for the determination of the numerical rank and the numerical null space,
there are alternative methods such as URV decomposition, LU decomposition, or QR factorization with column pivoting.
Now we can present our algorithm for computing the (m, n) reduced Padé approximant fm,n (z ) for f (z ) at the point z = a.
Algorithm.
Initialization:
f (z ) := the approximated function;
(m, n) := the order of the Padé approximant;
a := the center of expansion;
d := 0, if we want to delete vanishing coefficients of numerator and the denominator, else d := 1.
1. Compute the Taylor coefficients c0 , . . . , cm+n of f (z ) at the point z = a.
2. Form the Toeplitz matrix Tm = ∥ci−j+m ∥ i = 1, . . . , n + 1,
j = 1, . . . , n.
3. Determine its rank ~ = rank Tm and the index µ1 = m − n + ~ .
4. Form matrix Tµ1 +1 = ∥ci−j+µ1 +1 ∥ i = 1, . . . , m + n − µ1 ,
j = 1, . . . , n − m + µ1 + 1.
 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541 539

5. Test: Determine if the rank of Tµ1 +1 is equal to ~ . If not, interrupt calculation of the Padé approximant.
6. Find a basis for its one-dimensional kernel ker Tµ1 +1 .
The obtained vector (q0 , q1 , . . . , qn−m+µ1 )T is a vector formed from the coefficients of the minimal degree denominator
n−m+µ1
Q1 (z ) = qk (z − a)k of the Padé approximant.

k=0
7. If d = 0, then, for k = 0, . . . , ~ , do:
(k+1)
a. Form the matrix Tµ1 +1 .
(k+1)
b. Find its rank: rank Tµ1 +1 .
(k+1)
c. If rank Tµ1 +1 = ~ − 1, then replace qk in (q0 , q1 , . . . , qn−m+µ1 )T by zero.
Else step 6 should be omitted.
8. For k = 0, . . . , m, do pk = ck q0 + ck−1 q1 + · · · + ck−~ q~ and form the vector (p0 , . . . , pm )T consisting of the coefficients
m
of the numerator P1 (z ) = pk (z − a)k corresponding to Q1 (z ).

k=0
9. If d = 0, then, for k = 0, . . . , m, do:
[k]
a. Form the matrix Tµ1 +1 .
[k]
b. Find its rank: rank Tµ1 +1 .
c. If rank Tµ1 +1 = ~ , then replace pk in (p0 , . . . , pm )T by zero.
[k]

Else step 9 should be omitted.

End of Algorithm.
P (z )
Output: fm,n (z ) = Q1 (z ) .
1
Comments:
(1) Step 1, consisting of computing the Taylor coefficients, is very important for the algorithm. It is well known (see, for
example, [3]) that accuracy in the given coefficients ck is essential for Padé approximants. Usually most computing effort
goes into calculation of the coefficients rather than Padé approximants, and so the coefficients should be calculated with
the greatest possible accuracy. The way of calculating them depends essentially on the problem under consideration.
For example, in the Padé–Laplace method, coefficients ck are found from experimental data by numerical integration.
So the accuracy in the coefficients depends not only on the chosen method of integration but also on the quality of the
experiment.
(2) Step 5 is the control step. If the rank of Tµ1 +1 is not equal to ~ , then the computation of the Padé approximant should be
interrupted. The problem is most likely ill posed, and ranks are found wrongly. We face the problem of «not well-defined
numerical rank», or the tolerance tol is chosen improperly.
(3) In step 3, we find the essential index µ1 according to Proposition 3.3. The first essential polynomial, which is the minimal
degree denominator of the Padé approximant, is found in step 6 according to Theorem 3.1. Theorems 4.2 and 4.3 are
used in steps 7 and 9, respectively. The output approximant fm,n (z ) possesses properties listed in Theorem 4.1.
We have implemented the algorithm in procedure ReducedPade(f , m, n, a, d) in Maple (the code is available at
http://olga.6v6power.ru/soft), and now we would like to repeat Examples 2.3 and 2.4 from Section 2.
(z +1)(z −2)
Example 5.1. Let us find by ReducedPade the diagonal (4, 4) Padé approximant f4,4 for f (z ) = (z +2.1)(z −1) at the point z = 0
via the following commands.
(z +1)·(z −2)
f := (z +2.1)·(z −1) :
app := ReducedPade(f , 4, 4, 0, 0);
0.7773220744 + 0.3886610375z − 0.3886610371z 2
0.8161881781 − 0.427527140654618832z − 0.388661037357416250z 2
fsolve(numer (app), z , complex);
−0.9999999998, 2.000000001
fsolve(denom(app), z , complex);
−2.099999999, 1.000000000.

1.01
Example 5.2. Let us find the (2, 3) Padé approximant for the function f (z ) = (z +2z +
)(z −2.01)
at the point z = 0.
z +1.01
f := (z +2)·(z −2.01) :
app := ReducedPade(f , 2, 3, 0, 0);
−0.2438127455 − 0.2413987580z
0.9704230069 + 0.00241398781766979931z − 0.241398758054323787z 2
540 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541

fsolve(numer (app), z , complex);

−1.010000000
fsolve(denom(app), z , complex);
−1.999999999, 2.010000000.

As can be seen, the approximants obtained are identical to the approximated functions, and Froissart doublets do not
appear. Note that in the given examples the value of the parameter d was equal to zero. Hence, the vanishing coefficients of
the numerators and the denominators were deleted. The next example shows that such deletion is desirable.

z +1.01
Example 5.3. Let us find the (3, 7) Padé approximant for the function f (z ) = z 4 +3z 2 −4.01
at the point z = 0.
z +1.01
f := z 4 +3z 2 −4.01
:
app := ReducedPade(f , 3, 7, 0, 0);
0.1977728838 + 0.1958147364z
−0.7852170931 + 0.587444214531195330z 2 + 0.195814737672938278z 4
fsolve(numer (app), z , complex);
−1.010000000
fsolve(denom(app), z , complex);
−1.000999098, −2.000499738I , 2.000499738I , 1.000999098
app := ReducedPade(f , 3, 7, 0, 1);
(0.1977728838 + 0.1958147377z − 1.48862691 10−11 z 3 )/(−0.7852170931 − 5.23710724786852211 10−9 z
+ 0.587444214531194664z 2 − 1.13198619922094679 10−9 z 3 + 0.195814737672938056z 4
− 1.67404646568558580 10−10 z 5 )
fsolve(numer (app), z , complex);
−1.146906047 105 , −1.009999994, 1.146916147 105
fsolve(denom(app), z , complex);
−1.000999095, −1.728850777 10−9 − 2.000499738I , −1.728850777 10−9 + 2.000499738I ,
1.000999101, 1.169709095 109 .

6. Conclusion

This work was motivated by our intention of avoiding the appearance of Froissart doublets in signal processing using the
Padé–Laplace method. To do this, we have modified the classical definition of a Padé approximant by adding the requirement
of the minimal degree of its denominator. It turned out that this new definition of a reduced Padé approximant allows us to
avoid Froissart doublets induced by computer roundoff.
The reduced Padé approximant can be easily obtained by the proposed algorithm and always has nice properties which
the classical Padé approximant possesses only in the normal case: the denominator is unique up to a constant factor, the
numerator and the denominator do not have common non-zero roots, and their degrees and the deficiency index are known
exactly.

Acknowledgments

It is a pleasure to thank Professor A.L. Shestakov for his encouragement and guidance. We acknowledge the reviewer’s
comments and suggestions which were valuable in improving the quality of our manuscript. We are very grateful to L.N.
Trefethen, P. Gonnet, and S. Guttel for pointing out Gragg’s paper [6] to us and for useful comments on a preliminary version
of this paper that helped to improve the exposition.

References

[1] C. Brezinski, From numerical quadrature to Padé approximation, Appl. Numer. Math. 60 (2010) 1209–1220.
[2] S.P. Suetin, Padé approximants and efficient analytic continuation of a power series, Russian Math. Surveys 57 (1) (2002) 43–141.
[3] G.A. Baker Jr., P.R. Graves-Morris, Padé Approximants, second ed., Cambridge University Press, Cambridge, 1996.
 O.L. Ibryaeva, V.M. Adukov / Journal of Computational and Applied Mathematics 237 (2013) 529–541 541

[4] K.O. Geddes, Symbolic computation of Padé approximants, ACM Trans. Math. Softw. 5 (2) (1979) 218–233.
[5] A. Bultheel, Recursive algorithms for nonnormal Padé tables, SIAM J. Appl. Math. 39 (1) (1980) 106–118.
[6] W.B. Gragg, The Padé table and its relation to certain algorithms of numerical analysis, SIAM Rev. 14 (1) (1972) 1–62.
[7] J. Gilewicz, Y. Kryakin, Froissart doublets in Padé approximation in the case of polynomial noise, J. Comput. Appl. Math. 153 (2003) 235–242.
[8] P. Claverie, A. Denis, E. Yeramian, The representation of functions through the combined use of integral transforms and Padé approximants:
Padé–Laplace analysis of functions as sums of exponentials, Comput. Phys. Rep. 9 (1989) 247–299.
[9] V.M. Adukov, Generalized inversion of block Toeplitz matrices, Linear Algebra Appl. 274 (1998) 85–124.
[10] L.N. Trefethen, P. Gonnet, S. Guttel, Robust Padé approximation via SVD, SIAM Rev. (in press).
[11] G.L. Xu, A. Bultheel, Matrix Pade approximation: definitions and properties, Linear Algebra Appl. 137–138 (1990) 67–136.
[12] C. Pestano-Gabino, C. Gonzáles-Concepción, M.C. Gil-Fariña, A type of matrix Padé approximant inspired by scalar component models, J. Comput.
Appl. Math. 236 (2012) 3360–3372.
[13] G.H. Golub, C.F. Van Loan, Matrix Computations, third ed., JHU, Baltimore, MD, 1996.
[14] G.H. Golub, V. Klema, G.W. Stewart, Rank degeneracy and least squares problems, Tech. Rep. TR 456, 1976.
[15] L.V. Foster, Rank and null space calculations using matrix decomposition without column interchanges, Linear Algebra Appl. 74 (1986) 47–71.
[16] G.W. Stewart, Determining rank in the presence of errors, Tech. Rep. TR-92-108, 1992.
[17] O.L. Ibryaeva, A new algorithm for calculating Padé approximants and its Matlab implementation, in: Proc. of South Ural State University. Ser. of Math.
Simul. and Program., vol. 37, No. 254, 2011, pp. 99–107 (in Russian).